Are Prompt-based Models Clueless?

Proceedings of the 60th Annual Meeting of the Association for Computational Linguistics Volume 1: Long Papers, pages 2333 - 2352 May 22-27, 2022 c⃝2022 Association for Computational Linguistics Are Prompt-based Models Clueless? Pride Kavumba1,2 Ryo Takahashi3 1Tohoku University 2RIKEN AIP 3LegalForce Research kavumba.pride.q2@dc.tohoku.ac.jp {ryo.takahashi, yusuke.oda}@legalforce.co.jp Yasuke Oda3,1 Abstract Finetuning large pre-trained language models with a task-specific head has advanced the state- of-the-art on many natural language under- standing benchmarks. However, models with a task-specific head require a lot of training data, making them susceptible to learning and exploiting dataset-specific superficial cues that do not generalize to other datasets. Prompting has reduced the data requirement by reusing the language model head and formatting the task input to match the pre-training objective. Therefore, it is expected that few-shot prompt- based models do not exploit superficial cues. This paper presents an empirical examination of whether few-shot prompt-based models also exploit superficial cues. Analyzing few-shot prompt-based models on MNLI, SNLI, HANS, and COPA has revealed that prompt-based mod- els also exploit superficial cues. While the models perform well on instances with super- ficial cues, they often underperform or only marginally outperform random accuracy on in- stances without superficial cues. 1 Introduction Finetuning large pre-trained language models with a task-specific head has achieved remarkable performance on many natural language bench- marks (Wang et al., 2018, 2019). However, the task-specific head introduces a lot of random task- specific parameters that require enormous finetun- ing data to attain optimal performance. The expo- sure to enormous data increases the potential for models to learn and exploit dataset-specific super- ficial cues that do not generalize to other datasets without superficial cues (Gururangan et al., 2018; Poliak et al., 2018; Sugawara et al., 2018; Niven and Kao, 2019; Schuster et al., 2019; Kavumba et al., 2019). For example, Niven and Kao (2019) found that task-specific head models exploit the presence of “not” in the input of argument reason- ing comprehension dataset (Habernal et al., 2018) Figure 1: (A) shows a prompt-based model receiv- ing natural language inference (NLI) prompts gener- ated through a template. (B) Our analysis reveals that prompt-based models exploit superficial cues (high- lighted lexical overlap) that are predictive of entail- ment relation between premise and hypothesis. (C) The model fails to generalize to instances were the superfi- cial cues no longer predict entailment relation. to achieve state-of-the-art accuracy, but drop to random accuracy when the superficial cue is neu- tralized. On the other hand, prompting reuses the pre-training language model head, introducing no random task-specific parameters. Thus, prompt- based models can achieve remarkable performance with only a few training examples (Brown et al., 2020; Schick and Schütze, 2021a,b; Gao et al., 2021; Le Scao and Rush, 2021). Hence, few-shot prompting lowers the potential for models to learn and exploit dataset-specific superficial cues. This work empirically investigates whether few- shot prompt-based models exploit superficial cues. Specifically, we ask: Do few-shot prompt-based models exploit superficial cues? To answer this question, we examine prompted-based models on two fundamental tasks of …

Content truncated to 500 words